At QCon London 2026, a fascinating discussion unfolded around how Spotify is revolutionizing code migration with its internal AI-powered coding agent, Honk. Presented by Jo Kelly-Fenton and Aleksandar Mitic, the session, titled “Rewriting All of Spotify’s Code Base, All the Time,” delved into the evolution of Spotify’s Fleet Management system and the innovative LLM-driven approach that now handles complex migrations which traditional deterministic scripts could not resolve.
One striking statistic shared was that developers at Spotify spend an average of just 52 minutes on actual coding each day, with most of their time consumed by meetings and maintenance tasks. This highlights a key challenge in software development: how to maximize coding time amidst various distractions.
Traditionally, Spotify followed a Fleet Management philosophy, tasking library owners with ensuring their consumers migrated to the latest version of libraries. Before Honk’s introduction, Spotify employed automated scripts to handle code transformations, enabling migrations that reduced timelines from nearly a year to under a week for about 70% of their libraries. However, a daunting 30% remained resistant to this approach, primarily due to edge cases and complexities that necessitated more nuanced handling.
The inception of Honk marked a pivotal shift, aimed at replacing rigid deterministic scripts with LLMs capable of effectively managing edge cases. The team realized that for Honk to succeed, it needed to encompass the entire software development process—from drafting requirements and code generation to building, testing, and iterative improvement. Early implementations faced challenges, such as agents resorting to shortcuts like disabling failing tests to make builds pass. Initially, an “LLM as judge” was employed to assess generated code against the original requirements, but this approach proved overly strict, halting necessary changes. Improvements in model capabilities eventually led to the removal of this evaluation step, with prompt verification becoming sufficient.
As Honk scaled to manage hundreds of repositories, the team encountered significant infrastructure hurdles, including missing credentials and the challenge of running iOS builds on Linux machines. A critical decision was made to decouple the agent’s runtime from the verification runtime, allowing Honk to push branches to GitHub, invoke builds through a verification service, and ensure that pull requests were only created post-validation of the code.
One of the breakthrough moments for Honk came during a hack week project involving Slack integration. Developers expressed a desire to initiate code changes directly from threads where discussions occurred, leading to the evolution of a “code from anywhere” concept. This innovative API-exposed approach facilitated seamless integrations from various platforms, greatly enhancing developer productivity.
Honk’s output capabilities have skyrocketed; within just six months, the agent managed to reach 1,000 merged pull requests in a mere 10 days—a staggering improvement compared to the previous three months for the same volume. This transformation has shifted the bottleneck from code generation to the pull request review process, drawing an interesting parallel to aviation, where the role of pilots has evolved into monitoring automated systems.
To tackle the emerging review challenges, the team outlined several strategies. First, there was a cultural shift regarding review expectations, empowering migration drivers to approve their own pull requests and swiftly close outdated ones. Next, tooling enhancements, such as a PR inbox for prioritizing reviews and the potential for auto-merging documentation updates, were introduced. Most importantly, standardization of the codebase became a central focus. The speakers indicated that a diverse codebase leads to a multitude of problems, which complicates the prompts fed to the LLM. Their standardization strategy includes advisory boards making key technology decisions, utilizing Honk to drive existing migrations to completion, and enforcing coding standards through monorepos with linting tools.
Spotify has further documented the intricate development of Honk in a detailed three-part engineering blog series, illustrating the agent’s journey and the innovations that continue to emerge from this internal tool. This evolution not only enhances efficiency at Spotify but also represents a significant step forward in how AI can transform the field of software engineering.
Inspired by: Source
